CN109919921A - Based on the influence degree modeling method for generating confrontation network - Google Patents
Based on the influence degree modeling method for generating confrontation network Download PDFInfo
- Publication number
- CN109919921A CN109919921A CN201910143738.6A CN201910143738A CN109919921A CN 109919921 A CN109919921 A CN 109919921A CN 201910143738 A CN201910143738 A CN 201910143738A CN 109919921 A CN109919921 A CN 109919921A
- Authority
- CN
- China
- Prior art keywords
- model
- confrontation network
- data set
- image
- network
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Granted
Links
Landscapes
- Image Analysis (AREA)
Abstract
The invention belongs to image quality evaluation fields, a kind of image quality evaluation model that varying environment is different degrees of is constructed in turn to propose that a kind of application generates confrontation network technology binding hierarchy structural model, improve the accuracy of electric power picture recognition, the present invention, based on the influence degree modeling method for generating confrontation network, steps are as follows: the first step, carries out the collection of electric power image pattern;Second step, training generate confrontation network, and confrontation network is convolutional neural networks;Third step constructs hierarchy Model according to the data set that obtained generation confrontation network generates;4th step, analyzes experimental data, promotes the accuracy and high efficiency of appraisement system.Present invention is mainly applied to image quality evaluation occasions.
Description
Technical field
The invention belongs to image quality evaluation fields, are related to a kind of environment degree influence model using generation confrontation network
The method of foundation.
Background technique
With the rapid development of China's economic technology and the increasingly raising of public living standards, power consumer is to electricity
The demand of the power energy is increasingly vigorous, also higher and higher for the quality of power supply.Root of the power transmission terminal as electric system, play to
Important role is closed, the development of national electrical problem or even enterprise is directly related to.If large-scale power equipment misses
Difference will bring inestimable loss to national economy.Therefore, the safety problem of power transmission terminal is one of critical issue.
In recent years, the needs of the safety detection of power transmission terminal cause to be increasingly stringenter the condition of environment, rugged environment
The safe operation of the items of equipment of power transmission terminal will be directly affected.The complicated multiplicity of environment makes annual electric power accident have hair
It is raw, seriously threaten the safe operation of power grid.Natural weather influence is one of the main problem for jeopardizing power transmission network safety.How
Efficiently and the brought harm difficult to image detection of effective prevention natural weather is one of current urgent problem to be solved, right
It is played a crucial role in entire power grid normal operation.
Currently, the processing for electric power image pattern, there are mainly two types of situations.First is that the data sample influenced based on no environment
This is few, second is that being handled based on the more negative samples influenced with environment.Front and back is both limited by the weather environment shadow of outer bat
It rings, the largely electric power picture sample without environmental impact factor can not be obtained in the case of maximum probability.Make the acquisition of sample in this way
Cumbersome and time consuming, the result detected at the same time is difficult to ensure accuracy.With neural network, image recognition technology it is at full speed
Development, sample data required for being generated using particular technology have been applied in many fields, it is contemplated that image has at this stage
Many quality evaluation systems, it is contemplated that the different degrees of influence of varying environment, power transmission terminal are built at mountain plateau hilly country, day compression ring
Border is changeable, fights network technology using generating, and investigate varying environment influences the mechanism to image quality evaluation in various degree.
Generation confrontation network technology is the emerging technology of data sample needed for current generate, needed for being generated using generation network
Data sample, and arbiter then screens generation data, so that obtained sample set is more nearly true value[1]。
Based on image quality evaluation system dependent on generate confrontation network sample data and building and hierarchical structure mould
The foundation of type.It generates confrontation network and provides learning model end to end, the generation confrontation e-learning after training arrives image
Feature, can relatively improve and complete completed sample according to collection building.Compared to traditional image generating method, confrontation net is generated
Network has very outstanding feature extraction and feature learning ability, is had based on data set effect constructed by confrontation network is generated
Remarkable result is promoted[2]。
Therefore, it is utilized herein based on expansion research the problem of generating confrontation network varying environment Different Effects degree modeling
The negative sample of required a variety of a variety of degree of weather environment is generated based on the image generating technologies for generating confrontation network,
And then in the building for carrying out quality evaluation system by hierarchy Model, so as to reflect the quality of image from side.This
Sample can be this straight to ensure that the normal operation of electric power networks provides reliable guarantee with the building of the progress model of intelligent and high-efficiency
The appraisement system of sight has very important realistic meaning.
[1]Pan J,Ferrer C C,Mcguinness K,et al.SalGAN:Visual Saliency
Prediction with Generative Adversarial Networks[J].2017.
[2]Radford A,Metz L,Chintala S.Unsupervised Representation Learning
with Deep Convolutional Generative Adversarial Networks[J].Computer Science,
2015。
Summary of the invention
In order to overcome the deficiencies of the prior art, the present invention is directed to propose a kind of application generates confrontation network technology binding hierarchy knot
Structure model constructs a kind of image quality evaluation model that varying environment is different degrees of in turn, to greatly improve electric power picture recognition
Accuracy.For this reason, the technical scheme adopted by the present invention is that based on the influence degree modeling method for generating confrontation network,
Steps are as follows:
The first step carries out the collection of electric power image pattern, collects the picture of the power transmission terminal under various environment, and use image
Enhancing technology obtains being able to carry out training, being capable of satisfactory data set;
Second step, training generate confrontation network, and generating confrontation network essence is all convolutional neural networks, are obtained using step 1
The data set training arrived generates network, then by the data set and original data set training confrontation network by generating network generation, most
After reach Nash Equilibrium, i.e. confrontation network is beyond recognition the image for generating network;
Third step constructs hierarchy Model according to the data set that obtained generation confrontation network generates, according to step 2
Obtained last whole data set varying environment constructs hierarchy Model in various degree, and varying environment is divided into different etc.
Grade, evaluates the negative sample image for being input to this model, to establish an appraisement system;
4th step, analyzes experimental data, promotes the accuracy and high efficiency of appraisement system.
In second step further, a picture is randomly choosed from the data set of varying environment, and picture is converted to
Vector is denoted as x as true picture, and using x as the input for differentiating network, input value is one 0 to 1 after differentiating network
Between number, y be used for indicates input picture for true picture probability, be really 1, be generated as 0;
Loss function is calculated using obtained probability value, the picture type according to input is that generation image or true picture will
The label of the input data of discrimination model is labeled as 0 or 1;
The loss function of discrimination model:
-((1-y)log(1-D(G(z)))+ylogD(x)
Wherein G (z) is to generate model output, and D (x) is discrimination model output;Loss function as cross entropy loss function,
Loss is calculated, gradient anti-pass is carried out;
Generate the loss function of model:
(1-y)log(1-D(G(z)))
G (z) will obtain the data distribution of same original data set, therefore minimize generation model error, here produce G (z)
Raw error is transmitted to generation model, for the prediction result of discrimination model, is changed to the direction of change of gradient, if differentiating
Model thinks G (z) output when be real data set and thinks to export when be noise data, and gradient updating direction will be into
Row changes;
Final loss function are as follows:
WhereinThat is the prediction classification of discrimination model adjusts prediction probability, takes 0 to be used to change gradient direction, threshold value here
It is set as 0.5.
In third step further specifically, hierarchy Model is constructed;Construct pairwise comparison matrix;Mode of Level Simple Sequence and
Consistency check, that is, judge whether the pairwise comparison matrix of subjective building has preferable consistency on the whole;Total hierarchial sorting
And consistency check;
Construct the negative sample rating database generated based on electric power actual measurement sample by generating confrontation network, it is contemplated that this is negative
Sample data set weight is single, and the consistency of judgement is intuitive, so need to only consider that condition below establish can meet:
1. passing through abstract high-rise element, depression of order judgment matrix in hierarchical model;
2. being handled using the method that analytic hierarchy process (AHP) is combined with clustering methodology.
It is handled using the method that analytic hierarchy process (AHP) is combined with clustering methodology specifically, again using clustering methodology first
Using analytic hierarchy process (AHP), preferable evaluation result can be obtained, wherein the image of every kind of degree in each weather pattern is regarded as
One point, the connection of two different degrees of images in weather pattern of the same race is indicated using Euclidean distance, shows two apart from smaller
The degree of image is closer;Later by obtained cluster result, every layer of different weather difference is arranged in binding hierarchy analytic approach
The evaluation weight of degree finally determines overall model.
The features of the present invention and beneficial effect are:
The present invention devises a kind of electric power picture quality based on hierarchy Model using the algorithm for generating confrontation network
Appraisement system.The image influenced with natural weather environment clapped other than the appraisement system is research object, to obtained figure
As by generating confrontation network to obtain one including once again the different degrees of final data collection of varying environment, with this establish one it is complete
Kind and efficient appraisement system.
Detailed description of the invention:
Fig. 1 generates prototype network structure chart.
The different degrees of evaluation model of Fig. 2 electric power image different weather.
The original image that Fig. 3 is influenced without environment.
Five kinds of Fig. 4 different degrees of ambient images influenced with weather.
A, which by light rain is influenced image b and influenced image c by moderate rain, is influenced image by heavy rain
D, which is influenced image e by heavy rain, is influenced image by extra torrential rain.
Specific embodiment
The present invention constructs a kind of novel electric power picture appraisal model, it build by following steps and
Test.
The first step carries out the collection of electric power image pattern.The picture of the power transmission terminal under various environment is collected, and uses image
Enhancing technology obtains being able to carry out training, being capable of satisfactory data set.
Second step, training generate confrontation network.Generating confrontation network essence is all convolutional neural networks, is obtained using step 1
The data set training arrived generates network, then by the data set and original data set training confrontation network by generating network generation, most
After reach Nash Equilibrium, i.e. confrontation network is beyond recognition the image for generating network.
A picture is randomly choosed from the data set of varying environment, picture is converted to vector, as true picture, note
Make x.
Using x as differentiate network input, y as after differentiating network input value be one 0 to 1 between number, use
In indicating that input picture is the probability of true picture, it is really 1, is generated as 0.
Loss function is calculated using obtained probability value.Picture type according to input is that generation image or true picture will
The label of the input data of discrimination model is labeled as 0 or 1.
The loss function of discrimination model:
-((1-y)log(1-D(G(z)))+ylogD(x)
Loss function calculates loss as cross entropy loss function, carries out gradient anti-pass.
Generate the loss function of model:
(1-y)log(1-D(G(z)))
G (z) will obtain the data distribution of same original data set, therefore minimize generation model error, here produce G (z)
Raw error is transmitted to generation model.For the prediction result of discrimination model, the direction of change of gradient is changed.If differentiating
Model thinks G (z) output when be real data set and thinks to export when be noise data, and gradient updating direction will be into
Row changes.
Finally loss function is
WhereinThat is the prediction classification of discrimination model adjusts prediction probability, takes 0 to be used to change gradient direction, threshold value here
It is set as 0.5.
Third step constructs hierarchy Model according to the data set that obtained generation confrontation network generates.According to step 2
Obtained last whole data set varying environment constructs hierarchy Model in various degree.Varying environment is divided into different etc.
Grade, evaluates the negative sample image for being input to this model, to establish an appraisement system.
Construct hierarchy Model;Construct pairwise comparison matrix;Mode of Level Simple Sequence and consistency check (judge subjective structure
Whether the pairwise comparison matrix built has preferable consistency on the whole);(inspected layer is taken second place for total hierarchial sorting and consistency check
Between consistency).
Construct the negative sample rating database generated based on electric power actual measurement sample by generating confrontation network, it is contemplated that this is negative
Sample data set weight is single, and the consistency of judgement is intuitive, so need to only consider that condition below establish can meet.
The foundation of model is based on following two condition:
1. passing through abstract high-rise element, depression of order judgment matrix in hierarchical model
2. being handled using the method that analytic hierarchy process (AHP) is combined with clustering methodology
If first reusing analytic hierarchy process (AHP) using clustering methodology, preferable evaluation result can be obtained.Wherein by each day
The image of every kind of degree in gas type regards a point as, using Euclidean distance indicate in weather pattern of the same race two in various degree
The connection of image.It is closer apart from the smaller degree for showing two images, pass through obtained cluster result, binding hierarchy later
Analytic approach, the different degrees of evaluation weight of every layer of different weather of setting, finally determines overall model.
4th step, analyzes experimental data, promotes the accuracy and high efficiency of appraisement system.
For the elaboration technical solution being more clear, step will be illustrated in conjunction with specific structure.Specific mode by
Following step composition:
The first step prepares data set.
(1) preparing pictures data
The picture that the power transmission terminal acquired at power transmission terminal using machine is influenced with weather is collected, according in picture
Different weather conditions first divided, due to obtain early period with weather influence amount of images it is few, and influence degree without
Method distinguishes, and the geometric transformation for first passing through image increases the amount of image data, the influence institute based on environmental factor of the invention
Using degree of comparing transformation and noise disturbance.The range of data set has been expanded by the method that data enhance.
(2) image enhancement
Considered the small feature of data volume captured by power transmission terminal, in order to improve generate confrontation network accuracy and
Robustness carries out data enhancing to data set now.Herein from the Enhancement Method of five kinds of natural images to collected data set
Enhanced, wherein image of the S (i) as input terminal, S (o) is used as enhanced output image.
1) picture noise is considered first.Gauss salt-pepper noise and Gaussian noise both common picture noises are chosen,
In addition initial data set required for taking different signal-to-noise ratio to generate.Wherein N (θ) is noise, and θ is noise parameter.
S (o)=S (i)+N (θ)
2) image is blurred.The common filters such as mean filter, gaussian filtering and motion blur are chosen, are set different
It is filter that parameter, which obtains F (g) in the image-type after filter,As filter parameter.
3) histogram equalization.Converting probability density by integral probability density function for the histogram of original image is
The image of 1 (ideal situation), to have the function that improve contrast.The essence of histogram equalization is also a kind of specific region
Broadening, but will lead to whole image and converted to bright region.
According to the above several frequently seen scheme policies, first passes through and converted using above scheme, realize initial part number
Enhance according to the data of collection, preliminary data amount is 5 times of photographed data amount.
Second step, training generate confrontation network.
(1) the propagated forward stage
The present invention is used as error correction algorithms using common reversed passback, for updating model parameter.It generates model
Using full convolutional coding structure, Fig. 2 is the network architecture, and the number of the convolutional layer convolution kernel is indicated in the number in figure bracket.Network
In all convolution kernel size be 4 × 4, step-length is set as 2.Up-sampling and down-sampled proportionality factor are 2, i.e. convolutional layer
Half is reduced to characteristic image side length, warp lamination is enlarged into one times to characteristic image side length.The convolution of decoder the last layer
Layer uses Tanh function, and characteristic pattern is mapped as the output image of triple channel.
It is mostly that convolutional layer is followed by batch normalization layer that the convolutional layer that model and discrimination model use is generated in the present invention
(BatchNorm) structure of unit (Rectified Linear Units, ReLu) form is activated with nonlinear operation.It will
Dropout layers of Dropout rate are set as 50%.The ReLU used is set as 0.2 for LeakyReLu activation primitive, coefficient.
The convolutional layer of its last layer uses S sigmoid growth curve (Sigmoid) function, and characteristic pattern is mapped to one one
The output of dimension.
By data set obtained above, time random selection piece image pair, is sent into network and is trained every time.Training process
It is middle to set 1 for batchsize.
It is gradually fixed in training firstly, generating model, the weight of differentiation network is updated and identified with promoting it, divided
Class ability.Generator data set generated, which is put into arbiter, classifies, and is exactly the effect of arbiter, and arbiter will use up
It is possible to identify the picture for generating network, that is, it distinguishes the true from the false (synthesising picture).
Then, fixed to differentiate network, as soon as every generation weather environment influence factor figure in model is generated, by the composite diagram
It is distinguished in discriminator with the true weather environment figure with label, is to generate figure by the differentiation result of discrimination model
Error is back to generation model, more newly-generated Model Weight, so that generating the data that model generates can be closer to
Truthful data.
Next, fixed again generate model, arbiter training is carried out.By composite diagram compared with the original image based on label
It is trained, differentiates whether the image for generating each piece of image is true picture respectively, then take the average value of each piece of response to be used as and sentence
The final output of other model.Discrimination model, the weight of more newly-generated model are fixed again.
Repeatedly, the weight alternating iteration of fixed party more row another party, until both sides reach a dynamic equilibrium.This
When discrimination model which cannot be distinguished is the composite diagram for generating model and generating, it is believed that generate model produce it is close enough
Image is influenced like the environment of label.
Model is generated by the dual training repeatedly with discrimination model, makes great efforts to generate different from the varying environment based on label
The weather image of degree.Different every kind of weather are divided into five grades, generate enough sample datas, it can be by day compression ring
The characteristics of border, more significantly highlights, while the quality evaluation being also sufficient in various situations.
Third step constructs level mechanism model.
Fig. 2 show the different degrees of evaluation model of electric power image different weather.Here the first level is exactly data set itself.
Refer here to the final varying environment with label generated from generation confrontation network, different degrees of negative sample figure
Piece.It at this time can be mixed in together.Second level is the classification of different weather situation, enumerates three kinds of most common rain, snow, mist days here
Gas situation, actual conditions can also have comprising sand and dust, frost, hail etc., specific sample data appraisement system according to different weather
Body adjustment.Third layer be different weather in the case of different degrees of grade the case where, here only carried out five grades draw
Point, if, based on actual conditions, five grades have been able to meet big portion absolutely if necessary, the division of more multi-grade can be carried out
Grade separation in the case of point.Hierarchy Model needs to consider the problems of weight, therefore is set as phase for each layer here
Same weight, so as to not have to consider the influence of weight herein.
4th step, model measurement and recruitment evaluation.
After the completion of training, the image of test set is inputted into trained generation model, neural network forecast is obtained and goes out not
With the different degrees of image of environment.
By generating the observation of data set it can be found that the data set generated has five kinds under every kind of weather condition
Different degrees of weather condition.
Claims (4)
1. a kind of based on the influence degree modeling method for generating confrontation network, characterized in that steps are as follows:
The first step carries out the collection of electric power image pattern, collects the picture of the power transmission terminal under various environment, and use image enhancement
Technology obtains being able to carry out training, being capable of satisfactory data set;
Second step, training generate confrontation network, and generating confrontation network is convolutional neural networks, the data set obtained using step 1
Training generate network, then will by generate network generate data set and original data set training confrontation network, finally reach receive it is assorted
Equilibrium, i.e. confrontation network are beyond recognition the image for generating network;
Third step constructs hierarchy Model according to the data set that obtained generation confrontation network generates, according to obtained by step 2
To last whole data set varying environment construct hierarchy Model in various degree, varying environment is divided into different brackets, right
The negative sample image for being input to this model is evaluated, to establish an appraisement system;
4th step, analyzes experimental data, promotes the accuracy and high efficiency of appraisement system.
2. as described in claim 1 based on the influence degree modeling method for generating confrontation network, characterized in that second step
In further, a picture is randomly choosed from the data set of varying environment, picture is converted to vector, as true figure
Picture is denoted as x, and using x as the input for differentiating network, input value is the number between one 0 to 1 after differentiating network, and y is used for table
Show that input picture is the probability of true picture, is really 1, is generated as 0;
Loss function is calculated using obtained probability value, the picture type according to input is to generate image or true picture to differentiate
The label of the input data of model is labeled as 0 or 1;
The loss function of discrimination model:
-((1-y)log(1-D(G(z)))+ylogD(x)
Wherein G (z) is to generate model output, and D (x) is discrimination model output;Loss function is calculated as cross entropy loss function
Loss carries out gradient anti-pass;
Generate the loss function of model:
(1-y)log(1-D(G(z)))
G (z) will obtain the data distribution of same original data set, therefore minimize generation model error, here generate G (z)
Error is transmitted to generation model, for the prediction result of discrimination model, is changed to the direction of change of gradient, if discrimination model
Think that gradient updating direction will be changed when G (z) output is real data set and when thinking that output is noise data
Become;
Final loss function are as follows:
WhereinThat is the prediction classification of discrimination model adjusts prediction probability, takes 0 here for changing gradient direction, threshold value is set
It is 0.5.
3. as described in claim 1 based on the influence degree modeling method for generating confrontation network, characterized in that third step
In further specifically, construct hierarchy Model;Construct pairwise comparison matrix;Mode of Level Simple Sequence and consistency check, i.e.,
Judge whether the pairwise comparison matrix of subjective building has preferable consistency on the whole;Total hierarchial sorting and consistency check;
Construct the negative sample rating database generated based on electric power actual measurement sample by generating confrontation network, it is contemplated that this negative sample
Data set weight is single, and the consistency of judgement is intuitive, so need to only consider that condition below establish can meet:
1) pass through abstract high-rise element, depression of order judgment matrix in hierarchical model;
2) it is handled using the method that analytic hierarchy process (AHP) is combined with clustering methodology.
4. as described in claim 1 based on the influence degree modeling method for generating confrontation network, characterized in that use layer
The method that fractional analysis is combined with clustering methodology is handled specifically, first reuses step analysis using clustering methodology
Method can obtain preferable evaluation result, wherein regarding the image of every kind of degree in each weather pattern as a point, use Europe
Formula distance indicates the connection of two different degrees of images in weather pattern of the same race, more connects apart from the smaller degree for showing two images
Closely;Pass through obtained cluster result, binding hierarchy analytic approach, the different degrees of evaluation value of every layer of different weather of setting later
Weight finally determines overall model.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201910143738.6A CN109919921B (en) | 2019-02-25 | 2019-02-25 | Environmental impact degree modeling method based on generation countermeasure network |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201910143738.6A CN109919921B (en) | 2019-02-25 | 2019-02-25 | Environmental impact degree modeling method based on generation countermeasure network |
Publications (2)
Publication Number | Publication Date |
---|---|
CN109919921A true CN109919921A (en) | 2019-06-21 |
CN109919921B CN109919921B (en) | 2023-10-20 |
Family
ID=66962405
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201910143738.6A Active CN109919921B (en) | 2019-02-25 | 2019-02-25 | Environmental impact degree modeling method based on generation countermeasure network |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN109919921B (en) |
Cited By (4)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN110414362A (en) * | 2019-07-02 | 2019-11-05 | 安徽继远软件有限公司 | Electric power image data augmentation method based on production confrontation network |
CN110490659A (en) * | 2019-08-21 | 2019-11-22 | 广东电网有限责任公司 | A kind of customer charge curve generation method based on GAN |
CN111079642A (en) * | 2019-12-13 | 2020-04-28 | 国网浙江余姚市供电有限公司 | Line movable monitoring method and device and computer readable medium |
WO2021056843A1 (en) * | 2019-09-27 | 2021-04-01 | 北京市商汤科技开发有限公司 | Neural network training method and apparatus and image generation method and apparatus |
Citations (7)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN101334893A (en) * | 2008-08-01 | 2008-12-31 | 天津大学 | Fused image quality integrated evaluating method based on fuzzy neural network |
WO2016145571A1 (en) * | 2015-03-13 | 2016-09-22 | 深圳大学 | Method for blind image quality assessment based on conditional histogram codebook |
CN106952293A (en) * | 2016-12-26 | 2017-07-14 | 北京影谱科技股份有限公司 | A kind of method for tracking target based on nonparametric on-line talking |
CN108090902A (en) * | 2017-12-30 | 2018-05-29 | 中国传媒大学 | A kind of non-reference picture assessment method for encoding quality based on multiple dimensioned generation confrontation network |
CN108174225A (en) * | 2018-01-11 | 2018-06-15 | 上海交通大学 | Filter achieving method and system in coding and decoding video loop based on confrontation generation network |
CN108764298A (en) * | 2018-04-29 | 2018-11-06 | 天津大学 | Electric power image-context based on single classifier influences recognition methods |
CN108898269A (en) * | 2018-05-16 | 2018-11-27 | 天津大学 | Electric power image-context impact evaluation method based on measurement |
-
2019
- 2019-02-25 CN CN201910143738.6A patent/CN109919921B/en active Active
Patent Citations (7)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN101334893A (en) * | 2008-08-01 | 2008-12-31 | 天津大学 | Fused image quality integrated evaluating method based on fuzzy neural network |
WO2016145571A1 (en) * | 2015-03-13 | 2016-09-22 | 深圳大学 | Method for blind image quality assessment based on conditional histogram codebook |
CN106952293A (en) * | 2016-12-26 | 2017-07-14 | 北京影谱科技股份有限公司 | A kind of method for tracking target based on nonparametric on-line talking |
CN108090902A (en) * | 2017-12-30 | 2018-05-29 | 中国传媒大学 | A kind of non-reference picture assessment method for encoding quality based on multiple dimensioned generation confrontation network |
CN108174225A (en) * | 2018-01-11 | 2018-06-15 | 上海交通大学 | Filter achieving method and system in coding and decoding video loop based on confrontation generation network |
CN108764298A (en) * | 2018-04-29 | 2018-11-06 | 天津大学 | Electric power image-context based on single classifier influences recognition methods |
CN108898269A (en) * | 2018-05-16 | 2018-11-27 | 天津大学 | Electric power image-context impact evaluation method based on measurement |
Non-Patent Citations (3)
Title |
---|
PARIMALA KANCHARLA ET AL.: ""IMPROVING THE VISUAL QUALITY OF GENERATIVE ADVERSARIAL NETWORK (GAN)-GENERATED IMAGES USING THE MULTI-SCALE STRUCTURAL SIMILARITY INDEX"", 《IEEE》 * |
李方彪 等: ""生成式对抗神经网络的多帧红外图像超分辨率重建"", 《红外与激光工程》 * |
龙熙华 等: ""基于层次分析法的图像颜色传输效果评价"", 《计算机工程》 * |
Cited By (6)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN110414362A (en) * | 2019-07-02 | 2019-11-05 | 安徽继远软件有限公司 | Electric power image data augmentation method based on production confrontation network |
CN110490659A (en) * | 2019-08-21 | 2019-11-22 | 广东电网有限责任公司 | A kind of customer charge curve generation method based on GAN |
CN110490659B (en) * | 2019-08-21 | 2022-01-11 | 广东电网有限责任公司 | GAN-based user load curve generation method |
WO2021056843A1 (en) * | 2019-09-27 | 2021-04-01 | 北京市商汤科技开发有限公司 | Neural network training method and apparatus and image generation method and apparatus |
CN111079642A (en) * | 2019-12-13 | 2020-04-28 | 国网浙江余姚市供电有限公司 | Line movable monitoring method and device and computer readable medium |
CN111079642B (en) * | 2019-12-13 | 2023-11-14 | 国网浙江余姚市供电有限公司 | Line movable monitoring method and device and computer readable medium |
Also Published As
Publication number | Publication date |
---|---|
CN109919921B (en) | 2023-10-20 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN109919921A (en) | Based on the influence degree modeling method for generating confrontation network | |
CN107016405B (en) | A kind of pest image classification method based on classification prediction convolutional neural networks | |
CN109272500B (en) | Fabric classification method based on adaptive convolutional neural network | |
CN108229550B (en) | Cloud picture classification method based on multi-granularity cascade forest network | |
CN109934200A (en) | A kind of RGB color remote sensing images cloud detection method of optic and system based on improvement M-Net | |
CN106845556A (en) | A kind of fabric defect detection method based on convolutional neural networks | |
CN109460708A (en) | A kind of Forest fire image sample generating method based on generation confrontation network | |
CN110909690A (en) | Method for detecting occluded face image based on region generation | |
CN114612937B (en) | Pedestrian detection method based on single-mode enhancement by combining infrared light and visible light | |
CN102902956A (en) | Ground-based visible cloud image recognition processing method | |
CN110321959A (en) | A kind of coal rock detection method of multispectral image information and CNN | |
CN109948692A (en) | Picture detection method is generated based on the computer of multiple color spaces convolutional neural networks and random forest | |
CN108596243A (en) | The eye movement for watching figure and condition random field attentively based on classification watches figure prediction technique attentively | |
CN113297915A (en) | Insulator recognition target detection method based on unmanned aerial vehicle inspection | |
CN114255403A (en) | Optical remote sensing image data processing method and system based on deep learning | |
CN116152658A (en) | Forest fire smoke detection method based on domain countermeasure feature fusion network | |
CN104091327A (en) | Method and system for generating dendritic shrinkage porosity defect simulation image of casting | |
Ni et al. | Intelligent defect detection method of photovoltaic modules based on deep learning | |
CN112507881A (en) | sEMG signal classification method and system based on time convolution neural network | |
CN116309228A (en) | Method for converting visible light image into infrared image based on generation of countermeasure network | |
CN110490133A (en) | A method of children's photo being generated by parent's photo based on confrontation network is generated | |
CN117197554A (en) | Transformer oil leakage real-time detection method and system | |
CN110807372A (en) | Rapid optical remote sensing target identification method based on depth feature recombination | |
CN113435425B (en) | Wild animal emergence and emergence detection method based on recursive multi-feature fusion | |
CN115620050A (en) | Improved YOLOv5 aphid identification and counting method based on climate chamber environment |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
PB01 | Publication | ||
PB01 | Publication | ||
SE01 | Entry into force of request for substantive examination | ||
SE01 | Entry into force of request for substantive examination | ||
GR01 | Patent grant | ||
GR01 | Patent grant |